Biological Conservation 125 (2005) 459–465 www.elsevier.com/locate/biocon 0006-3207/$ - see front matter  2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.biocon.2005.04.018 Using self-organizing maps to investigate spatial patterns of non-native species Régis Céréghino ¤ , Frédéric Santoul, Arthur Compin, Sylvain Mastrorillo Laboratoire d’Ecologie des Hydrosystèmes, UMR 5177 CNRS/UPS, Université Paul Sabatier, 118 route de Narbonne, F-31062 Toulouse cedex 4, France Received 15 November 2004 Available online 8 June 2005 Abstract Our ability to demonstrate statistical patterns of invasion by non-native species will determine the success of future management projects. We investigated the suitability of self-organizing maps (SOM, neural network) for patterning habitat invasion by exotic Wsh species at the regional scale (Southwest France), using a binary dataset of species occurrences. The SOM visualization can be used as an analytical tool to bring out relationships between sample locations and biological variables, but in addition the weight of each species in the output of the SOM can be interpreted as its occurrence probability in various geographic areas. After training the SOM with Wsh presence/absence data, the k-means algorithm helped to derive three major clusters of sites (headwater, montane, and plain areas). Each cluster was divided into two subsets of sites according to non-native Wsh, because assemblage compositions delineated diVerent geological areas: Pyrenees Mountains, Massif Central Mountains, and alluvial plain. Occurrence probabilities of species within our study stream system were roughly inXuenced by river typology, with a higher occurrence probability for most species (i.e. a greater risk) in downstream sections. Conversely, headwater streams at the highest elevations in the study area showed the lowest risk of invasion. EYcient analytical tools such as SOM may thus help to yield explicit schemes that could inXuence the measures to be taken in the latter phase of conservation plans.  2005 Elsevier Ltd. All rights reserved. Keywords: Neural networks; Biological invasions; Fish; Stream system; Bioassessment 1. Introduction Biological invasions and the presence of exotic species are a pervasive and costly environmental problem (Vito- usek et al., 1996; Lake and Leishman, 2004), that has been the focus of intense management activities over the last decades (Mooney and Drake, 1986). However, pre- dicting habitat susceptibility to invasion remains a diY- cult task, and the obtained results are often elusive (Kennedy et al., 2002). In order to prevent invasions and target monitoring eVorts more eVectively, we thus need to forecast locations at the greatest risk of invasions (Leung et al., 2004; Holway, 2005). Fish species were often studied as model organisms under this topic (Koehn, 2004), certainly due to Wshing and/or economic concerns. To allow predictive models to work across a wide range of conditions, it is important to examine pat- terns associated with Wsh invasions at broad spatial scales (e.g., a region, a large stream system) (Gido et al., 2004). Moreover, there is a need to develop alternative analytical approaches which can maximize the informa- tion extracted from available data, such as “simple” presence/absence data (Bessa-Gomes and Petrucci-Fons- eca, 2003). Inspired by the structure and the mechanism of the human brain, artiWcial neural networks (ANNs) provide ¤ Corresponding author. Tel.: +33 561 558 436; fax: +33 561 556 096. E-mail address: cereghin@cict.fr (R. Céréghino).